Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Wed, 11 Aug 2021 09:12:36 -0400
From:   Peter Xu <peterx@redhat.com>
To:     Mingwei Zhang <mizhang@google.com>
Cc:     Jim Mattson <jmattson@google.com>,
        Paolo Bonzini <pbonzini@redhat.com>,
        Sean Christopherson <seanjc@google.com>,
        Vitaly Kuznetsov <vkuznets@redhat.com>,
        Wanpeng Li <wanpengli@tencent.com>,
        Joerg Roedel <joro@8bytes.org>, kvm <kvm@vger.kernel.org>,
        LKML <linux-kernel@vger.kernel.org>,
        Ben Gardon <bgardon@google.com>,
        David Matlack <dmatlack@google.com>,
        Jing Zhang <jingzhangos@google.com>
Subject: Re: [PATCH v4 3/3] KVM: x86/mmu: Add detailed page size stats
Message-ID: <YRPMxLdL5vsZRyux@t490s>
References: <20210803044607.599629-1-mizhang@google.com>
 <20210803044607.599629-4-mizhang@google.com>
 <CALMp9eR4AB0O7__VwGNbnm-99Pp7fNssr8XGHpq+6Pwkpo_oAw@mail.gmail.com>
 <CAL715WJLfJc3dnLUcMRY=wNPX0XDuL9WzF-4VWMdS_OudShXHg@mail.gmail.com>
 <CAL715WLO9+CpNa4ZQX4J2OdyqOBsX0+g0M4bNe+A+6FVxB2OxA@mail.gmail.com>
 <YRMKPd2ZarXCX6vm@t490s>
 <CAL715WJWPzBqmjeTJ6mZa=dUaF5+MdqaCrk5CEzvcz1X99cm0g@mail.gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
In-Reply-To: <CAL715WJWPzBqmjeTJ6mZa=dUaF5+MdqaCrk5CEzvcz1X99cm0g@mail.gmail.com>
Precedence: bulk

On Tue, Aug 10, 2021 at 06:06:51PM -0700, Mingwei Zhang wrote:
> Regarding the pursuit for accuracy, I think there might be several
> reasons. One of the most critical reasons that I know is that we need
> to ensure dirty logging works correctly, i.e., when dirty logging is
> enabled, all huge pages (both 2MB and 1GB) _are_ gone. Hope that
> clarifies a little bit?

It's just for statistics, right?  I mean dirty log should be working even
without this change.

But I didn't read closely last night, so we want to have "how many huge pages
we're mapping", not "how many we've mapped in the history".  Yes that makes
sense to be accurate.  I should have looked more carefully, sorry.

PS: it turns out atomic is not that expensive as I thought even on a 200 core
system, which takes 7ns (but for sure it's still expensive than normal memory
ops, and bus locking); I thought it'll be bigger as on a 40 core system I got
15ns which is 2x of my laptop of 8 cores, but it didn't really grow but shrink.

Thanks,

-- 
Peter Xu